Usage

• unit-test.ipynb is the main logic to 1. train DDPM with EWC and 2. analysis Fisher information matrix.
• You want to first train DDPM on task 0, before applying EWC and training on task 1
• We have three versions of Fisher used in EWC: diagnoal, rank-1 and rank-1 optimal. compute_rank1_coeff_and_mean function computes all necessary values.
  • mu is the rank-1 vector
• You want to compute the Fisher by inferencing on the old task
• The test FID block is the main evaluation metric. Say the model is sequentially trained on task 0 and task 1. We want to evaluate the model's generation quality on task 0 and 1. Here we compare against the test set.
• To emprically analysis the full Fisher information matrix, we need a small DDPM by uncommentting the configurations in the model initialization step.